Lumen GI and Reflections feedback thread

But how can we get some good quality reflections and refraction? Raytracing seems still better. In any case, still didn’t find the solution to my problem yet.

i think true raytraced refraction is a lil durther down the line. i can bounce that code thru my mind’s eye. raytracing thru translucent surfaces and correctly blending their colors over the next bounce is not that easy. it’s a “reverse” blend. a ray payload that gotta be carried thru for every translucent refraction bounce and a ton of math at the end. the color accumulation per bounce is the issue. it’s not very efficient.

imo… lumen reflections are as good they can be for now, while still maintaining interactive framerates on a broad spectrum of hardware. so… ima say… deal with it. shrugs

1 Like

Ok and good to know, but whatever quality number I use they’re always too blurred!
On the sphere’s reflection you can also see that the reflected shadows are too precise and there is no “area shadows” effect.



you sure you use full 0 roughness on that ball? also… i noticed that you use post chromatism. how much does that do to the reflection? i don’t wanna build a test rn.

my test reflections are clean af. ofc temporal in motion. i don’t mind that. no aggressive post volumes either.

screenshot is a out of the box 5.5 daily, btw. build it 3 hours ago. looks good. my highrise map got some nice fps now. not stable but doesn’t dip below 40 anymore at 100% 1080. i mean… i’m totally rt limited in that one, anyway. my gpu runs it without sweating. :slight_smile:

1 Like

had to notice the sun size broke in my map. a lil test was in order. directionality. source angle or soft angle?

the soft angle does some overcast looking reflections but hard shadows.

using both does overcast shadows, but the reflection is lost. or… added on top.

i’d guess a kinda fusion would be nice looking for this sun. :slight_smile: it’s to late for me to start coding, tho. hmm

They trimmed the lumen RT payload size to something like a third of what it was in UE4. I’m honestly pretty impressed that they got refraction working well at all given that constraint. I wonder if substrate will result in some sort of payload rearchitecting to avoid all the cost.

Thanks and I’ll check that!
As for the post chromatism, I dont’think it can change the blurry feeling on the reflections too much, but I need to check this too.

Not tested, but probably Lumen reflections are only reflecting hard shadows. RT reflections could reflect hard shadows and area shadows, if selected. Not sure if Lumen is compatible or if it has the same switch, to select the kind of reflected shadows.

Have you tried disabling temporal accumulation, as commented before?

Hey! It is a great expierence to use it.

2 Likes

Lumen reflections have had a performance increase in 5.4 by 1ms in a test scenario at 1440p on a desktop 3060!

I’m talking legit 1ms, not .2 under or over. That 1ms is pretty awesome, but TSR also become more expensive by 1ms(again no .2 under or over) which canceled out any performance gain Lumen got but I don’t even like TSR so it’s not a problem for me.

Btw, this is the first time UE5.4 compiled so I’ll update this post regarding anymore issues(hoping to see Lumen remains stable in motion without TAA etc)

EDIT: Feedback:

  • A Cvar that would really help things out would to limit reconstruction from Bilateral and Reconstruction samples from affecting lower roughness like 0.0, 0.01. Reconstruction is needed and more beneficial at 0.05-0.4. When close to 0.0 or 0.01, ruin what temporal accumulation already does a good job for.

  • Lumen Reflections do not jitter in a way where 0.0 reflections can be constructed to the main internal resolution as frames pass on low roughness, if they were, we could get high res reflections in around 5 frames. I guess this was unfortunately handed off to TSR.

To see what I mean, Mirror reflections with (r.Lumen.Reflections.DownsampleFactor 2 + r.Lumen.Reflections.ScreenSpaceReconstruction 0) looks the same as: (r.Lumen.Reflections.DownsampleFactor 1 +r.Lumen.Reflections.ScreenSpaceReconstruction 1)

r.Lumen.Reflections.Temporal.MaxRayDirections 32 is really great at 1/4 res reflections the only thing missing is standalone temporal reconstruction on 0.0-0.05.

5.3 r.Lumen.Reflections.ScreenSpaceReconstruction 0 takes off 5.55555556 of reflection cost vs 22.3214286% in 5.4. In 5.3 Downsample factor 4 takes off 48.71% vs 50% in 5.4

EDIT: The following images have no reflections enabled. In 5.4 Lumen Probe Gather doesn’t give as much specular compared to 5.3(in 5.3, Lumen reflections appear pink in the same area on the blue cube thing with just probe gather) but, this does make it more like pathtracing



1 Like

Interesting things. Just to point: I think you are already using the release candidate of the 5.4, as the branch is closed since 4 days ago, so I think this will be the Preview release for the Launcher.

1 Like

Just a quick first test:

In Epic quality, differences in GI are huge:

UE5.3:

UE5.4 Preview 1:


(Yeah, a little performance drop in 5.4, but in the other hand reflections are slightly better).

When changing booth to Cinematic, they look almost the same, in the right way.

Also noticed that Lumen with r.Lumen.Reflections.Temporal = 0 is much more worse in 5.4. (And yes, disabling temporal accumulation for reflections can be an actual use-case, as it can achieve better results, just combining it with other cvars and parameters to reduce the noise. In fact, if I’m not wrong, we will need to disable it if we want to use DLSS ray reconstruction). First, UE5.3; secondly, UE5.4:

Off-topic: the edge’s artifacts are still there, as nobody knows who is the responsible of this area of development (RT shadows):

1 Like

If you can post a stat gpu shot, interested in seeing where the cost is coming from. Also, kinda looks splotchy on the curtain and we still have lighting leaking.

If I’m not mistaken, traces are done against radiance probes? I feel like this should start including world normal information to prevent this. Ofc, that might only work for bake-able Lumen.

1 Like

The couch and curtain look broken. Maybe the cloth shading model isn’t working right?

The difference is indeed big, but in favor of 5.3, not 5.4. 5.4 looks just worse in this comparison. Could you do a path tracer render to see which one is closer to it?

2 Likes

Hi @TheKJ , @ZacD , @Rawalanche , it was a quick test because I’m travelling and won’t have much time to further tests until the end of the month, sorry.

Anyway, I did a quick check: the main difference between 5.3 and 5.4 was due to having a value of 0.8 (or less) in Final gather quality in post process volume, instead of 1 (anyway, the difference shouldn’t be so huge for a delta of 0.2). When 1 or higher, they look very similar, as intended.

5.4, in overall, looks more stable in GI and in reflections too, but a (very) little less performant as tradeoff. Less splotches in general, but they are stronger, I think (like using a bigger sampler).

Here was a small test. The project was stress test for all features made in 5.3, then with a bookmark I measured performance. Then copied the project(small in MB, large scene) and switch it to 5.4, same place, same resolution. Not a pretty scene but the performance speaks volumes. Most of the scene is below .4 roughness, very little quad overdraw, high (“60fps”) settings.


Could you do a path tracer render to see which one is closer to it?

did that in this post, but I can go ahead and do an archviz scene soon.

EDIT: Gezz, just realized Lumen GI doubled in cost(both had it in async by default).

lil performance screenshare. 5.4 preview. stable 48 fps with all the bounces. i didn’t hit that in the daily with the floor reflection enabled. lights spammed. shadows softened. no dlss ofc. engine firing on all cylinders. this is great.

edit: nvm. i shot at 720p internal. my stupid blueprint, again. xP either way… the good features are in. and it’s absolutely usable to make good looking games above 30 fps (i can smooth sync 36 to my panel).

Hey! :slight_smile:

I saw your post in the 5.4 Preview thread and wanted to let you know that we’ve already logged this issue internally. This actually stems from Ray Traced Shadows - From a previous fix on the shadows side, we get better shadows but also reveals this artifacting. For tracking of the issue, please see UE-200307 once available.

3 Likes

I see, so to make sure, you are looking here for a three state switch: [auto, two-sided, one-sided] instead of a current checkbox “Generate as two-sided”?

That’s an interesting point. We have something similar in reflections, where reconstruction and bilateral filters are scaled by roughness, and temporal accumulation is reduced for mirror like surfaces. I’ve made a note to look into it to see if it can be improved.

We already have jitter from screen pixels and not sure what else could be jittered for roughness=0.

4 Likes